VainF / DeepLabV3Plus-Pytorch

Pretrained DeepLabv3 and DeepLabv3+ for Pascal VOC & Cityscapes

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

New additional classes not training

CowbungaMk1 opened this issue · comments

Hello, I am working on a project that facilitates adding new classes to the cityscapes dataset and adding these classes to dataset/cityscapes.py. However, the training doesn't seem to ever learn these classes and the IOU is always 'nan' for 19-22 (the added classes). If i switch the class ids, such as my new class 22 to 0 (in place of road) it will train that class just fine but not for the other class. Is there some weight initialization step I'm missing? thanks

Hello @CowbungaMk1, please modify this line to add new classes.

Thanks for your quick response. I have already tried that and it still trains the original classes just fine. However the additional classes always return a MIOU of 'nan' to visdom and don't appear when using predict.

Thanks

issues resolved